1 ; RUN: opt < %s -slsr -gvn -S | FileCheck %s 2 3 target datalayout = "e-i64:64-v16:16-v32:32-n16:32:64-p:64:64:64-p1:32:32:32" 4 5 ; foo(input[0]); 6 ; foo(input[s]); 7 ; foo(input[s * 2]); 8 ; => 9 ; p0 = &input[0]; 10 ; foo(*p); 11 ; p1 = p0 + s; 12 ; foo(*p1); 13 ; p2 = p1 + s; 14 ; foo(*p2); 15 define void @slsr_gep(i32* %input, i64 %s) { 16 ; CHECK-LABEL: @slsr_gep( 17 ; v0 = input[0]; 18 %p0 = getelementptr inbounds i32, i32* %input, i64 0 19 call void @foo(i32* %p0) 20 21 ; v1 = input[s]; 22 %p1 = getelementptr inbounds i32, i32* %input, i64 %s 23 ; CHECK: %p1 = getelementptr inbounds i32, i32* %input, i64 %s 24 call void @foo(i32* %p1) 25 26 ; v2 = input[s * 2]; 27 %s2 = shl nsw i64 %s, 1 28 %p2 = getelementptr inbounds i32, i32* %input, i64 %s2 29 ; CHECK: %p2 = getelementptr inbounds i32, i32* %p1, i64 %s 30 call void @foo(i32* %p2) 31 32 ret void 33 } 34 35 ; foo(input[0]); 36 ; foo(input[(long)s]); 37 ; foo(input[(long)(s * 2)]); 38 ; => 39 ; p0 = &input[0]; 40 ; foo(*p); 41 ; p1 = p0 + (long)s; 42 ; foo(*p1); 43 ; p2 = p1 + (long)s; 44 ; foo(*p2); 45 define void @slsr_gep_sext(i32* %input, i32 %s) { 46 ; CHECK-LABEL: @slsr_gep_sext( 47 ; v0 = input[0]; 48 %p0 = getelementptr inbounds i32, i32* %input, i64 0 49 call void @foo(i32* %p0) 50 51 ; v1 = input[s]; 52 %t = sext i32 %s to i64 53 %p1 = getelementptr inbounds i32, i32* %input, i64 %t 54 ; CHECK: %p1 = getelementptr inbounds i32, i32* %input, i64 %t 55 call void @foo(i32* %p1) 56 57 ; v2 = input[s * 2]; 58 %s2 = shl nsw i32 %s, 1 59 %t2 = sext i32 %s2 to i64 60 %p2 = getelementptr inbounds i32, i32* %input, i64 %t2 61 ; CHECK: %p2 = getelementptr inbounds i32, i32* %p1, i64 %t 62 call void @foo(i32* %p2) 63 64 ret void 65 } 66 67 ; int input[10][5]; 68 ; foo(input[s][t]); 69 ; foo(input[s * 2][t]); 70 ; foo(input[s * 3][t]); 71 ; => 72 ; p0 = &input[s][t]; 73 ; foo(*p0); 74 ; p1 = p0 + 5s; 75 ; foo(*p1); 76 ; p2 = p1 + 5s; 77 ; foo(*p2); 78 define void @slsr_gep_2d([10 x [5 x i32]]* %input, i64 %s, i64 %t) { 79 ; CHECK-LABEL: @slsr_gep_2d( 80 ; v0 = input[s][t]; 81 %p0 = getelementptr inbounds [10 x [5 x i32]], [10 x [5 x i32]]* %input, i64 0, i64 %s, i64 %t 82 call void @foo(i32* %p0) 83 84 ; v1 = input[s * 2][t]; 85 %s2 = shl nsw i64 %s, 1 86 ; CHECK: [[BUMP:%[a-zA-Z0-9]+]] = mul i64 %s, 5 87 %p1 = getelementptr inbounds [10 x [5 x i32]], [10 x [5 x i32]]* %input, i64 0, i64 %s2, i64 %t 88 ; CHECK: %p1 = getelementptr inbounds i32, i32* %p0, i64 [[BUMP]] 89 call void @foo(i32* %p1) 90 91 ; v3 = input[s * 3][t]; 92 %s3 = mul nsw i64 %s, 3 93 %p2 = getelementptr inbounds [10 x [5 x i32]], [10 x [5 x i32]]* %input, i64 0, i64 %s3, i64 %t 94 ; CHECK: %p2 = getelementptr inbounds i32, i32* %p1, i64 [[BUMP]] 95 call void @foo(i32* %p2) 96 97 ret void 98 } 99 100 %struct.S = type <{ i64, i32 }> 101 102 ; In this case, the bump 103 ; = (char *)&input[s * 2][t].f1 - (char *)&input[s][t].f1 104 ; = 60 * s 105 ; which may not be divisible by typeof(input[s][t].f1) = 8. Therefore, we 106 ; rewrite the candidates using byte offset instead of index offset as in 107 ; @slsr_gep_2d. 108 define void @slsr_gep_uglygep([10 x [5 x %struct.S]]* %input, i64 %s, i64 %t) { 109 ; CHECK-LABEL: @slsr_gep_uglygep( 110 ; v0 = input[s][t].f1; 111 %p0 = getelementptr inbounds [10 x [5 x %struct.S]], [10 x [5 x %struct.S]]* %input, i64 0, i64 %s, i64 %t, i32 0 112 call void @bar(i64* %p0) 113 114 ; v1 = input[s * 2][t].f1; 115 %s2 = shl nsw i64 %s, 1 116 ; CHECK: [[BUMP:%[a-zA-Z0-9]+]] = mul i64 %s, 60 117 %p1 = getelementptr inbounds [10 x [5 x %struct.S]], [10 x [5 x %struct.S]]* %input, i64 0, i64 %s2, i64 %t, i32 0 118 ; CHECK: getelementptr inbounds i8, i8* %{{[0-9]+}}, i64 [[BUMP]] 119 call void @bar(i64* %p1) 120 121 ; v2 = input[s * 3][t].f1; 122 %s3 = mul nsw i64 %s, 3 123 %p2 = getelementptr inbounds [10 x [5 x %struct.S]], [10 x [5 x %struct.S]]* %input, i64 0, i64 %s3, i64 %t, i32 0 124 ; CHECK: getelementptr inbounds i8, i8* %{{[0-9]+}}, i64 [[BUMP]] 125 call void @bar(i64* %p2) 126 127 ret void 128 } 129 130 define void @slsr_out_of_bounds_gep(i32* %input, i32 %s) { 131 ; CHECK-LABEL: @slsr_out_of_bounds_gep( 132 ; v0 = input[0]; 133 %p0 = getelementptr i32, i32* %input, i64 0 134 call void @foo(i32* %p0) 135 136 ; v1 = input[(long)s]; 137 %t = sext i32 %s to i64 138 %p1 = getelementptr i32, i32* %input, i64 %t 139 ; CHECK: %p1 = getelementptr i32, i32* %input, i64 %t 140 call void @foo(i32* %p1) 141 142 ; v2 = input[(long)(s * 2)]; 143 %s2 = shl nsw i32 %s, 1 144 %t2 = sext i32 %s2 to i64 145 %p2 = getelementptr i32, i32* %input, i64 %t2 146 ; CHECK: %p2 = getelementptr i32, i32* %p1, i64 %t 147 call void @foo(i32* %p2) 148 149 ret void 150 } 151 152 define void @slsr_gep_128bit_index(i32* %input, i128 %s) { 153 ; CHECK-LABEL: @slsr_gep_128bit_index( 154 ; p0 = &input[0] 155 %p0 = getelementptr inbounds i32, i32* %input, i128 0 156 call void @foo(i32* %p0) 157 158 ; p1 = &input[s << 125] 159 %s125 = shl nsw i128 %s, 125 160 %p1 = getelementptr inbounds i32, i32* %input, i128 %s125 161 ; CHECK: %p1 = getelementptr inbounds i32, i32* %input, i128 %s125 162 call void @foo(i32* %p1) 163 164 ; p2 = &input[s << 126] 165 %s126 = shl nsw i128 %s, 126 166 %p2 = getelementptr inbounds i32, i32* %input, i128 %s126 167 ; CHECK: %p2 = getelementptr inbounds i32, i32* %input, i128 %s126 168 call void @foo(i32* %p2) 169 170 ret void 171 } 172 173 define void @slsr_gep_32bit_pointer(i32 addrspace(1)* %input, i64 %s) { 174 ; CHECK-LABEL: @slsr_gep_32bit_pointer( 175 ; p1 = &input[s] 176 %p1 = getelementptr inbounds i32, i32 addrspace(1)* %input, i64 %s 177 call void @baz(i32 addrspace(1)* %p1) 178 179 ; p2 = &input[s * 2] 180 %s2 = mul nsw i64 %s, 2 181 %p2 = getelementptr inbounds i32, i32 addrspace(1)* %input, i64 %s2 182 ; %s2 is wider than the pointer size of addrspace(1), so do not factor it. 183 ; CHECK: %p2 = getelementptr inbounds i32, i32 addrspace(1)* %input, i64 %s2 184 call void @baz(i32 addrspace(1)* %p2) 185 186 ret void 187 } 188 189 declare void @foo(i32*) 190 declare void @bar(i64*) 191 declare void @baz(i32 addrspace(1)*) 192
